1 / 36

Differential (De)Serialization for Optimized SOAP Performance

Differential (De)Serialization for Optimized SOAP Performance. Michael J. Lewis Grid Computing Research Laboratory Department of Computer Science Binghamton University State University of New York (with Nayef Abu-Ghazaleh, Madhu Govindaraju ). Motivation.

tory
Download Presentation

Differential (De)Serialization for Optimized SOAP Performance

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Differential (De)Serialization for Optimized SOAP Performance Michael J. Lewis Grid Computing Research Laboratory Department of Computer Science Binghamton University State University of New York (with Nayef Abu-Ghazaleh, Madhu Govindaraju)

  2. Motivation • SOAP is an XML-based protocol for Web Services that (usually) runs over HTTP • Advantages • extensible, language and platform independent, simple, robust, expressive, and interoperable • The adoption of Web Services standards for Grid computing requires high performance

  3. The SOAP Bottleneck Serialization and deserialization • The in memory representation for data must be converted to ASCII and embedded within XML • Serialization and deserialization conversion routines can account for 90% of end-to-end time for a SOAP RPC call [HPDC 2002, Chiu et. al.] • Our approach • Avoid serialization and deserialization altogether, whenever possible • bSOAP: Binghamton’s SOAP implementation

  4. Overview of the Optimizations • Differential Serialization (DS) (sender side) • Save a copy of the last outgoing message • If the next call’s message would be similar, then • use the previous message as a template • only serialize the differences from the last message • Differential Deserialization (DDS) (receiver side) • Checkpoint incoming message portions • If the next incoming message is similar, then • use the deserialized values from the last message • only deserialize the differences from the last message

  5. DS and DDS • DS and DDS are separate, different, disjoint optimization techniques • sender side (DS) vs. receiver side (DDS) • data update tracking (DS) vs. parser checkpointing (DDS) • neither depends on the other • each takes advantage of sequences of similar messages to avoid expensive SOAP message processing • neither changes the protocol, what goes in the SOAP message, or on the wire • each remains interoperable with other SOAP implementations

  6. DS: Update Tracking • How do we know if the data in the next message will be the same as in the previous one? • If it is different, how do we know which parts must be reserialized? • How can we ensure that reserialization of message parts does not corrupt other portions of the message?

  7. Data Update Tracking (DUT) Table struct MIO { int a; int b; double val;}; int mioArray(MIO[] mios) Field TPointer SLength FWidth Dirty? X 5 5 YES Y3 7 YES Z 5 10 NO POST /mioExample HTTP/1.1 . <?xml version='1.0'?><SOAP-ENV:Envelope ...> . <x xsi:type='xsd:int'>12345</x> <y xsi:type='xsd:int'>678</y> <z xsi:type='xsd:double'>1.166</val> . </SOAP-ENV:Envelope>

  8. Problems and Approaches • Problems • Some fields require reserialization • The current field width may be too small for the next value • The current message (or chunk) size may be too small Solving these problems enables DS, but incurs overhead • Approaches • shifting • chunking • stuffing • stealing • chunk overlaying

  9. Shifting • Shifting: Expand the message on-the-fly when the serialized form of a new value exceeds its field width • Shift the bytes of the template message to make room • Update DUT table entries for all shifted data …</w><x xsi:type='xsd:int'>1.2</x><y xsi:type=…. becomes …</w><x xsi:type='xsd:int'>1.23456</x><y xsi:type=…. • Performance penalty • DUT table updating, memory moves, possible memory reallocation

  10. Stuffing • Stuffing: Allocate more space than necessary for a data element • explicitly when the template is first created, or after serializing a value that requires less space • Helps avoid shifting altogether • Doesn’t work for strings, base64 encoding …<y xsi:type='xsd:int'>678</y><z xsi:type=… can be represented as …<y xsi:type='xsd:int'>678</y><z xsi:type=…

  11. Stealing • Stealing: Take space from nearby stuffed fields • Can be less costly than shifting [ISWS ‘04] …'>678</y><z xsi:type='xsd:double'>1.166</val> y can steal from z to yield… …'>677.345</y><z xsi:type='xsd:double'>1.166</val> • Performance depends on several factors • Halting Criteria: When to stop stealing? • Direction: Left, right, or back-and-forth?

  12. Performance • Performance depends on • which techniques are invoked • “how different” the next message is • Message Content Matches • identical messages, no dirty bits • Perfect Structural Matches • data elements and their sizes persist • Partial Structural Matches • some data elements change size • requires shifting, stealing, stuffing, etc. • We study the performance of all our techniques on synthetic workloads of scientific data • summary: 17%  10X improvement

  13. Perfect Structural Matches • Perfect Structural Matches: • Some data items must be overwritten (DUT table dirty bits) • No shifting required • Performance study: • vary the message size • vary the reserialization percentage • vary the data type • Doubles and • Message Interface Objects (MIO’s, <int, int, double>) (not shown)

  14. Send Time depends directly on % serialized • Important to avoid reserializing

  15. DDS: The Approach • As an incoming message is being processed • store parser states periodically • compute corresponding message portion checksums • For subsequent (hopefully similar) messages • compare incoming message checksums with stored checksums (fast mode parsing) • if checksums match • the parser can skip to the next parser state without actually generating it from the incoming message • on checksum mismatch • revert to regular mode parsing

  16. Effectiveness • Depends on… • Similarity in consecutive messages • determines how often in fast mode • How much faster fast mode is • deserialization vs. checksum calc and compare • Efficiency in identifying mode switches • Checkpoint and checksum overhead

  17. Creating Checkpoints • First one right after start tag that contains the name of the back end element • Thereafter, checkpoints are created periodically • based on number of bytes processed • configurable parameter of bSOAP • Tradeoff: overhead vs. fast mode processing time • standard implementation: full parser checkpoints • optimization [Grid 2005]: differential checkpoints

  18. Fast mode parsing • Parser reads messages and computes checksums on message portion boundaries • a match allows the parser to “skip” to the next saved state • Switching back to fast mode • must compare current and stored parser states • matching stack contains necessary structural info • namespace aliases must also be the same • stored and checked separately

  19. Performance Summary • Without DDS: comparable to gSOAP • DDS Overhead • message portions 256  4K; overhead < 10% • message portion size 32 bytes too small • DDS improvement • large “hard to deserialize” messages, very similar  25X speedup (upper bound) • Dual mode performance • depends on message portion size, where and how often mode switches take place • can reduce by a factor of 3, or be slightly slower

  20. Benchmark Suite for SOAP-Based Grid Web Services • Motivation • Web services based applications have diverse requirements • SOAP and XML present design and implementation challenges • Several novel efforts exist to address key bottlenecks • examples can be found in gSOAP, bSOAP, .NET • A benchmark suite • can help determine the best available toolkit • based on communication patterns and data structures in use • Benchmarks and performance evaluation framework • Drivers, WSDL files and Java code • helps provide insights on opportunities for optimization • Madhu Govindaraju: mgovinda at cs.binghamton.edu • http:// grid.cs.binghamton.edu / projects / soap_bench

  21. Thank You • For More information • Grid Computing Research Laboratory • SUNY – Binghamton: Computer Science Department • http:// grid.cs.binghamton.edu • mlewis at binghamton.edu • DS: [HPDC ’04], [IC ‘04] • DDS: [SC ’05], [Grid2005]

  22. Extra Slides

  23. Experimental Setup • Machines • Dual Pentium 4 Xeon 2.0 GHz, 1 GB DDR RAM, 15K RPM 18 GB Ultra-160 SCSI drive. • Network • Gigabit Ethernet. • OS • Debian Linux. Kernel version 2.4.24. • SOAP implementations • bSOAP and gSOAP v2.4 compiled with gcc version 2.95.4, flags: -O2 • XSOAP 1.2.28-RC1 compiled with JDK 1.4.2 • bSOAP/gSOAP socket options: SO_KEEPALIVE, TCP_NODELAY,SO_SNDBUF = SO_RCVBUF = 32768 • Dummy SOAP Server (no deserialization).

  24. Message Content Matches • Message Content Match: • The entire stored message template can be reused without change • No dirty bits in the DUT table • Best case performance improvement • Performance Study • compare gSOAP, XSOAP, and bSOAP, with differential serialization on and off • vary the message size • vary the data type: doubles and MIO’s (not shown)

  25. bSOAP ~= gSOAP • 10X imprvmt in DS • (expected result) • Upper bound

  26. Shifting • Partial Structural Match: • Not all of array elements are reserialized • Performance Study • Intermediate size values to maximum size values. • Array of doubles (18  24) • Array of MIO’s (36 46)(not shown)

  27. 100%  75%: Imprvt 23% 75%  50%: Imprvt 31% 50%  25%: Imprvt 46%

  28. Stuffing • Closing Tag Shift: • Stuffed whitespace comes after the closing tag • Must move the tag to accommodate smaller values • Performance Study • send smallest values (1 char) • vary field size: smallest, intermediate, maximum • Array of doubles (max = 24, intermediate = 18, min = 1) • Array of MIOs • (max = 46, intermediate = 38, min = 3) (not shown)

  29. Closing tag shift, not increased message size, effects stuffing performance

  30. Summary • SOAP performance is poor, due to serialization and deserialization • Differential serialization • Save a copy of outgoing messages, and serialize changes only, to avoid the observed SOAP bottleneck • Techniques: • Shifting, chunking, chunk padding, stuffing, stealing, chunk overlaying • Performance is promising (17% to 10X improvement), depends on similarity of messages

  31. Other Approaches • SOAP performance improvements • Compression • Base-64 encoding • External encoding: Attachments (SwA), DIME • These approaches may be necessary and can be effective. However • they undermine SOAP’s beneficial characteristics • interoperability suffers • The goal • improve performance, retain SOAP’s benefits

  32. Applications that can Benefit • Differential Serialization is only beneficial for applications that repeatedly resend similar messages • Such applications do exist: • Linear system analyzers • Resource information dissemination systems • Google & Amazon query responses • etc.

  33. Data Update Tracking (DUT) Table • Each saved message has its own DUT table • Each data element in the message has its own DUT table entry, which contains: • Location: A pointer to the data item’s current location in the template message • Type: A pointer to a data structure that contains information about the data item's type. • Serialized Length: The number of characters needed to store the last written value • Field Width: The number of allocated characters in the template • A Dirty Bit indicates whether the data item has been changed since the template value was written

  34. Updating the DUT Table • DUT table dirty bits must be updated whenever in-memory data changes • Current implementation • explicit programmer calls whenever data changes • Eventual intended implementation • more automatic • variables are registered with our bSOAP library • data will have accessor functions through which changes must be made • when data is written, the DUT table dirty bits can be updated accordingly • disallows “back door” pointer-based updates • requires calling the client stub with the same input param variables

  35. Worst Case Shifting • “Worst case shifting”: • All values are reserialized from smallest size values to largest size values. • Performance Study • vary the chunk size (8K and 32K) • Array of doubles (1  24). • Array of MIOs (3  46) (not shown)

  36. Worst case shifting is 4X slower Reducing chunk size doesn’t help

More Related